The application of queuing network models to computer systems

Abstract

The evaluation and prediction of computer systems is difficult because of the complexity of today's modern computer systems. It is therefore important that effective techniques and tools are available for computer performance evaluation. This thesis describes a performance study of a real computer system using one such technique, analytic modelling, which is based on queuing theory. The class of solvable queuing networks and the approach in deriving their performance measures are described. Approximate solution methods applicable to network models with no analytical results are also examined, including cases where exact solutions are less attractive because of the high computational cost. Operational analysis Introduced by Buzen in 1976 to solve some queue networks is also described. The main contribution of this thesis is in modelling an actual computer system, an IBM 4341 operating under the virtual machine concept. Measurement tools on the system - INDICATE, SMART, VM/MONITOR and its data reduction facility, VMAP, were used where appropriate to estimate the data required by the model and for testing and validating the predictions. The model was solved using both operational analysis and the results of Baskett, Chandy, Muntz and Palacios. For a single homogeneous class of users, the results were found to be identical.